A GraphSAGE-based model with fingerprints only to predict drug-drug interactions

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence

Background:

• Combination drug therapy improves efficacy for complex diseases but risks adverse effects.
• Accurate prediction of drug-drug interactions (DDIs) is crucial for safe and effective treatment.
• Existing deep learning models for DDI prediction often require extensive drug property data, limiting their applicability.

Purpose of the Study:

• To develop a novel deep learning-based model for predicting DDIs.
• To design a model that utilizes commonly available drug fingerprint features for wider applicability.
• To improve the accuracy and reliability of DDI prediction compared to existing methods.

Main Methods:

• Drugs were represented using common fingerprint features.
• A graph convolutional network method, GraphSAGE, was employed to fuse fingerprint features with the drug interaction network, generating high-level drug features.
• The inner product was used to score the strength of potential drug pairs.

Main Results:

• The model achieved high performance with an AUROC of 0.9704 and AUPR of 0.9727 via 10-fold cross-validation.
• Performance surpassed models using only fingerprint features and was competitive with models using more comprehensive drug properties.
• Ablation tests confirmed the significance of the model's components, and analysis revealed its strengths and limitations across drugs with varying network degrees.

Conclusions:

• The developed model effectively predicts DDIs using accessible drug fingerprint features and graph convolutional networks.
• Identified novel DDIs suggest potential therapeutic benefits (e.g., PEA and cannabinol) and risks (e.g., WIN 55,212-2 and cannabinol).
• This approach provides valuable insights for optimizing combination therapies and mitigating adverse drug events.